Object tracking using the trust-region method in scale-space with probability distribution images generated by kernel density estimation

In the widely used mean shift-based tracking algorithms, targets are described by color histograms with their size determined using predefined parameters. However histograms affect the precision of the target's color description and it's impossible to precisely describe a target's size with discrete parameters. This paper presents a new approach for tracking objects in image sequences which precisely tracks the constant changes of the target's size. A reference image gives the kernel density estimate (KDE) of the target's color distribution. Then for each incoming frame, a probability distribution image of the target is created through evaluating and normalizing of the KDE. Through searching the local maxima of multi-scale normalized Laplacian filters of the probability distribution image the target is located and its size is determined. Comparison with histogram-based algorithms show that the new method describes the target more accurately and thus achieves much better tracking precision.